KR20200063904A - Method and apparatus for measuring relevance between nodes of edge-labeled multigraph - Google Patents

Abstract

A method for measuring relevance between nodes in a graph having edges with labels representing relations between nodes comprises the steps of: learning a rule in which a label of a surfer changes as the surfer moves between nodes included in the graph; normalizing an adjacency matrix of the graph for each edge label; and using a result of learning the rule and the normalized adjacency matrix to calculate a score matrix through a repetitive technique.

Description

Method and apparatus for measuring relevance between vertices in a multi-edge label graph {METHOD AND APPARATUS FOR MEASURING RELEVANCE BETWEEN NODES OF EDGE-LABELED MULTIGRAPH}

Embodiments disclosed herein relates to a method and apparatus for measuring a relevance between vertices by reflecting a label in a graph in which a trunk is labeled with a relationship between vertices.

Recently, the importance of a technique for measuring relevance between vertices included in a graph for link prediction, recommendation system, anomaly detection, and community detection on an online network is increasing.

The RWR (Random Walk with Restart) technique is widely used as a method for measuring the relevance between vertices in a graph consisting of vertices and edges, which measures the relevance between vertices based on the structure of the graph.

Many of the graphs actually used are given a label indicating the relationship between the vertices on the edge, and according to the RWR technique, only the structure of the graph is considered, so that the label is not reflected when measuring the relevance between vertices.

That is, according to the existing RWR technique, it is possible to find out only that the two vertices are related to each other, and there is a limitation that it is not possible to grasp the specific relationship.

On the other hand, the above-mentioned background technology is the technical information acquired by the inventor for the derivation of the present invention or acquired in the derivation process of the present invention, and is not necessarily a known technology disclosed to the general public before filing the present invention. .

In the embodiments disclosed in the present specification, in measuring a relevance between vertices in a graph in which a label indicating the relationship between vertices is applied to the trunk, by reflecting the label assigned to the trunk, it is possible to grasp the relationship between the vertices It is intended to provide a method and apparatus.

As a technical means for achieving the above-described technical problem, according to an embodiment, in a graph where a trunk is labeled with a relationship between vertices, a method of measuring the degree of relevance between vertices is a vertex included in the graph Learning a rule in which the label of the surfer changes as the surfer moves between the fields, normalizing the adjacency matrix of the graph for each edge label, and learning the rule and the normalized adjacency. It may include using a matrix and calculating a score matrix through an iterative technique.

According to another embodiment, a computer program for performing a method of measuring a degree of relevance between vertices in a graph on which a trunk is labeled with a relationship between vertices, the method of measuring relevance included in the graph Learning a rule in which the label of the surfer changes as the surfer moves between the vertices to be generated, normalizing the adjacency matrix of the graph for each edge label, and learning the rule and normalizing the result. And calculating a score matrix through an iterative technique.

According to another embodiment, in a graph labeled trunks showing relationships between vertices, a computer readable recording medium in which a program for performing a method of measuring relevance between vertices is recorded. The method includes learning a rule in which a surfer's label changes as the surfer moves between vertices included in the graph, normalizing the adjacency matrix of the graph for each edge label, and the rule It may include using the result of learning and the normalized neighbor matrix, and calculating a score matrix through an iterative technique.

According to another embodiment, in a graph in which a trunk is labeled with a relationship between vertices, an apparatus for measuring the degree of relevance between vertices is a communication unit for receiving information about a graph that is a target of relevance measurement , A storage unit for storing a program for measuring the degree of relevance between vertices of the graph, and a control unit for measuring the degree of relevance between vertices in the graph by executing the program, wherein the control unit includes vertices included in the graph As the surfer moves between them, learn the rules for changing the label of the surfer, normalize the adjacency matrix of the graph for each edge label, and learn the result of learning the rule and the normalized adjacency matrix. The score matrix can be computed using and iterative techniques.

According to any one of the above-described problem solving means, when measuring the relevance between vertices of a graph, a label indicating the relationship between the two vertices connected through each edge is reflected, so that the You can expect the effect of figuring out whether there is a relationship.

The effects obtained in the disclosed embodiments are not limited to the above-mentioned effects, and other effects not mentioned are obvious to those skilled in the art to which the embodiments disclosed from the following description belong. Can be understood.

1 is a view showing an example of calculating the relevance between vertices by applying an RWR technique to a graph.
2 is a diagram illustrating an apparatus for measuring a degree of relevance between vertices of a graph according to an embodiment.
3 is a flow chart for explaining a method for measuring a relationship between vertices of a graph according to an embodiment.
4 and 5 are diagrams for explaining a process of learning a rule in which a label of a surfer changes as a surfer moves in a graph labeled with a trunk according to an embodiment.
6 is a flowchart illustrating a process of learning a rule in which a label of a surfer changes as a surfer moves in a graph labeled with a trunk according to an embodiment.
7 is a flowchart illustrating a process of normalizing an adjacent matrix in which a label of a graph is reflected according to an embodiment.
8 is a flowchart illustrating a process of calculating a score matrix through an iterative technique according to an embodiment.

Hereinafter, various embodiments will be described in detail with reference to the accompanying drawings. The embodiments described below may be embodied in various different forms. In order to more clearly describe the features of the embodiments, detailed descriptions of the items well known to those of ordinary skill in the art to which the following embodiments pertain are omitted. In the drawings, parts irrelevant to the description of the embodiments are omitted, and like reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a component is "connected" to another component, this includes not only a case of'directly connecting', but also a case of'connecting other components in between'. In addition, when a configuration is said to "include" a configuration, this means that, unless specifically stated otherwise, it may mean that other configurations may be included as well.

Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings.

In the present specification, in a graph in which a label indicating a relationship between vertices is assigned to an edge line, a labeled random walk with restart (LRWR) technique for measuring relevance between vertices by reflecting a label is introduced. , Embodiments in which the LRWR technique is utilized will be described. When using the LRWR technique, it is possible to infer the custom between any two vertices by using the information in the graph.

Since the LRWR technique is based on the RWR (Random Walk with Restart) technique, which is a method of measuring relevance between representative vertices, the RWR technique will be described first.

Graph mining refers to the task of discovering useful information in a graph. In addition, the graph ranking technique is one of the techniques for performing graph mining, and is a technique for determining the importance of vertices in a graph composed of vertices (edges) and edges (edges) or proximity between vertices. As a graph ranking technique, a RWR technique based on a random walk is often used, and the RWR technique will be described in detail with reference to FIG. 1.

1 is a view showing an example of calculating the relevance between vertices by applying an RWR technique to a graph. 1 is a graph consisting of twelve vertices and trunks connecting the vertices. Each vertex is assigned a vertex number from 1 to 12. The RWR technique is a random surfer (1), which is a virtual person moving from one starting vertex to one starting vertex (or seed node) among a plurality of vertices included in the graph. (Hereafter, a surfer) is a technique to calculate the probability of reaching other vertices as a score. The score (RWR score) calculated by the RWR technique indicates the degree of relevance between the starting and other vertices, and the higher the score, the higher the relevance between the two vertices.

In FIG. 1, vertex 4 was selected as the starting vertex. The surfer 1 starts at the starting vertex 4 and performs either random walk or restart action. For example, the surfer 1 may perform an action (random walk) that moves to any one of vertices 1, 3, or 5, or may perform an action (restart) to return to vertex 4. In addition, if the surfer 1 has reached the vertex 9 after several movements, the next action is to perform an action (random walk) to move to one of the vertices 8 or 10, or return to the vertex 4 You can perform an action (restart).

That is, each time the surfer 1 moves once, it immediately moves to the neighboring vertex (random walk) or moves to the starting vertex (restart). At this time, if the probability that the surfer 1 performs a restart is c, the probability of performing a random walk is (1-c).

The probability that the surfer 1 reaches each vertex is called an RWR score, and a column vector r consisting of an RWR score for each vertex is called an RWR score vector. Also, if the adjacency matrix representing the graph is A, the RWR score vector r can be expressed by the following equation.

는 각 행의 원소의 합이 1이 되도록 정규화된 인접 행렬이다. 수학식 1로부터 다음의 수학식 2를 얻을 수 있다.At this time, q is a column vector in which the value of the element corresponding to the starting vertex is 1 and the remaining elements are all 0. In addition,

Is an adjacency matrix normalized so that the sum of the elements in each row is 1. From Equation 1, the following Equation 2 can be obtained.

를 H라고 한다면, RWR 점수 벡터 r은 다음의 수학식 3과 같은 선형 시스템의 해로 구할 수 있다.therefore,

If H is H, the RWR score vector r can be obtained by solving the linear system as in Equation 3 below.

There are two methods for obtaining RWR score vector r: iterative approach and preprocessing approach. Each technique will be described below.

First, the iterative technique is a method in which Equation 1 is repeatedly applied to an arbitrary initial value for r so that the value of r converges within a certain range. Therefore, the iterative technique is expressed by the following equation (4).

는 허용 오차 (error tolerance) 또는 수렴 문턱값이라 한다.At this time, i means the number of times the calculation is repeatedly performed. Starting from the initial value r ⁽⁰⁾ , Equation 4 can be repeatedly applied until Equation 5 below is satisfied. Of course, Equation 5 is an example of a convergence condition, and it is also possible to set the convergence condition differently. At this time,

Is called error tolerance or convergence threshold.

This time, the pretreatment technique will be described. In the pre-processing technique, most calculations for obtaining the RWR score are performed in advance, and when the starting vertex is selected, the selected starting vertex is substituted into the pre-prepared result. Specifically, the following Equation 6 can be obtained from Equation 3, and thus, if the inverse matrix of H is obtained, the RWR score vector r can be quickly obtained even if the starting vertex is changed.

The LRWR technique described below is based on the iterative technique of the RWR technique described above.

The LRWR technique presents a method for effectively grasping the relationship between two vertices when there is a label in the graph consisting of the vertices and the edge showing the relationship between the vertices. That is, the purpose of the LRWR technique is to reflect the label of the trunk in the relevance score between vertices.

To this end, the LRWR technique requires that 1) the surfer has a label indicating the relationship between the starting vertex and the currently visited vertex, and 2) that the surfer's label is in accordance with the label of the trunk included in the path the surfer travels while performing a random walk. By making changes, multi-hop relational reasoning is enabled, and 3) learn from the graph how the surfer's label changes as the surfer moves to make correct inferences.

2 is a diagram illustrating an apparatus for measuring a degree of relevance between vertices of a graph according to an embodiment. The apparatus 100 for measuring relevance according to an embodiment may include a communication unit 110, a control unit 120, and a storage unit 130.

The communication unit 110 is a configuration for performing communication with a network such as the Internet, and may be configured with a communication chipset supporting Ethernet communication and wireless LAN communication. The relevance measurement apparatus 100 may receive information on a graph that is a target of relevance measurement from the outside through the communication unit 110.

The control unit 120 is a configuration including at least one processor such as a CPU, and controls overall operation of the relevance measurement device 100. In particular, the control unit 120 measures the relevance between vertices of the graph by applying the LRWR technique by executing the program stored in the storage unit 130. A detailed method for the controller 120 to measure the relevance between vertices of the graph will be described with reference to FIGS. 3 to 8 below.

Various types of programs and data may be stored in the storage unit 130. In particular, a program for measuring the degree of relevance between vertices of a graph by applying an LRWR technique may be stored in the storage unit 130 and executed by the control unit 120.

Hereinafter, with reference to FIGS. 3 to 8, a process in which the relevance measuring apparatus 100 measures the relevance between vertices from the graph labeled with the trunk will be described in detail.

3 is a flow chart for explaining a method for measuring a relationship between vertices of a graph according to an embodiment.

Referring to FIG. 3, the controller 120 of the relevance measuring apparatus 100 learns a rule in which the label of the surfer changes as the surfer moves between vertices included in the graph in step 301. A detailed process of performing step 301 will be described below with reference to FIGS. 4 to 6.

In step 302, the controller 120 normalizes the adjacency matrix of the graph for each label. A detailed process of performing step 302 will be described below with reference to FIG. 7.

In step 303, the controller 120 uses the learning result of step 301 and the normalized adjacency matrix in step 302, and calculates a score matrix through an iterative technique. A detailed process of performing step 303 will be described below with reference to FIG. 8.

First, in step 301, a process in which the controller 120 learns a rule in which the label of the surfer is changed will be described in detail with reference to FIGS. 4 to 6.

4 and 5 are diagrams for explaining a process of learning a rule in which a label of a surfer changes as a surfer moves in a graph labeled with a trunk according to an embodiment.

The graph shown in FIG. 4 is assigned to any one of the labels “childOf”, “grandchildOf” and “spouseOf” to the edges. The surfer does not have a label at the starting vertex s, but has a label while moving to another vertex along the edge, and the surfer's label can change as the surfer moves.

In the graph shown in FIG. 4, it is assumed that there are two rules as follows.

Rule 1: When a surfer with the label of “childOf” moves along the edge with the label of “childOf”, the label of the surfer is changed to “grandchildOf”.

Rule 2: When a surfer with the label “grandchildOf” moves along the edge labeled “spouseOf”, the surfer's label does not change.

Looking at a path in FIG. 4(a) where the surfer moves from the starting vertex (source node, s) to the target vertex (target node, t), the surfer does not have a label at the starting vertex s, but a “childOf” label. If you move to vertex u along the edge, you will have the label "childOf".

Also, if the surfer moves from vertex u to vertex v, he has the label “grandchildOf” according to rule 1.

In addition, when the surfer moves from vertex v to target vertex t, the label of “grandchildOf” is maintained according to rule 2.

Therefore, it can be seen that there is a relationship of “grandchildOf” between the starting vertex s and the target vertex t.

The label of the surfer who has moved through multi-hop relational reasoning can be determined according to the determined rule. Therefore, if learning the rule in which the label of the surfer is changed, the label of the trunk is placed on the path of the surfer. Can reflect.

Meanwhile, the two rules assumed above may be derived from a label transitive triangle as shown in (b) and (c) of FIG. 4.

To find out the rules for multi-hop relationship inference as described above, explaining how to use the data-driven approach to extract the knowledge contained in the graph by extracting the label transition triangle from the given graph is as follows. same.

이라는 라벨 전이 관찰은 라벨

를 갖는 서퍼가 라벨

의 간선을 따라 이동하면 서퍼의 라벨이

로 변경됨을 의미한다.From the label transition triangle shown in Fig. 5A, the label transition observation shown in Fig. 5C can be derived. At this time,

The label metastasis observation is labeled

Surfer having a label

As you move along the trunk of the

Means change to

The specific process of deriving the label transition observation from the label transition triangle is as follows.

를 갖는 간선을 따라 정점 y로 이동하면 서퍼는 라벨

를 갖게되고, 이와 유사하게 서퍼가 정점 x로부터 라벨

를 갖는 간선을 따라 정점 z로 이동하면 서퍼는 라벨

를 갖게된다. 따라서, 라벨

를 갖는 서퍼가 라벨

를 갖는 간선을 따라 이동하면 서퍼의 라벨이

로 변경된다는 라벨 전이 관찰을 도 5의 (a)에 도시된 라벨 전이 삼각형으로부터 도출할 수 있다.As can be seen in Fig. 5(b), a surfer without a label is labeled from vertex x.

When you move to vertex y along the edge with the surfer labels

And similarly the surfer labels from vertex x

When you move to vertex z along the edge with the surfer labels

Will have Therefore, the label

Surfer having a label

As you move along the trunk with the label of the surfer

The label transition observation that is changed to can be derived from the label transition triangle shown in Fig. 5A.

On the other hand, it is possible to obtain a rule in which the label of the surfer is changed from the observation of the label transition, and the process is as follows.

First, all label transition triangles are extracted from the graph using a triangle enumeration algorithm, and label transition observations are derived from the extracted label transition triangles.

에 기반하여 최대 가중 우도 추정(Maximum Weighted Likelihood Estimation, 이하 MWLE)을 적용하여 라벨 전이 확률 행렬(label transition probability matrix)

를 산출할 수 있다. Label transfer observation and label weight

Label transition probability matrix by applying maximum weighted likelihood estimation (MWLE) based on

Can be calculated.

는 라벨

에 대한 라벨 전이 확률 행렬이며,

의

번째 원소인

의 값은

이다. 이때,

는 라벨 전이 확률(label transition probability)로서 라벨

를 갖는 서퍼가 라벨

의 간선을 따라 이동할 경우 서퍼의 라벨이

로 변경될 확률을 의미한다. 따라서, 다음의 수학식 7과 같은 관계가 성립된다.

Label

Is the label transition probability matrix for

of

The second element

The value of

to be. At this time,

Is the label transition probability

Surfer having a label

When moving along the trunk of the

Means the probability of change to. Therefore, the relationship as shown in the following equation (7) is established.

를 학습하기 위해 다음과 같은 과정을 거친다.Label transition probability to maximize likelihood function making label transition observations

In order to learn, it goes through the following process.

세트가 주어진다고 가정한다.First, the label transition observations expressed by Equation 8 below

Assume that a set is given.

는

에 포함된 라벨 전이 관찰들의 개수이고,

는 라벨 전이 관찰

을 의미한다. 또한,

는 시작 정점에서의 서퍼의 라벨이고,

는 목표 정점에서의 서퍼의 라벨이다. 즉,

는 서퍼가 시작 정점에서 목표 정점까지 이동했을 때 레벨

를 갖게 되는 라벨 전이 관찰을 의미한다.At this time,

The

The number of label metastasis observations included in

Observe label metastasis

Means In addition,

Is the surfer's label at the starting vertex,

Is the surfer's label at the target vertex. In other words,

Level when the surfer moves from the starting vertex to the target vertex

It means to observe the label metastasis that will have.

는 파라미터

와 관련된

의 확률이라고 하면, 로그-우도 함수(log-likelihood function)

는 다음의 수학식 9와 같이 표현된다.Meanwhile,

Is the parameter

Related to

Speaking of probability, log-likelihood function

Is expressed by the following equation (9).

의 값을 최대화하는 것은 만족스럽지 않은 결과를 가져올 수 있다.However, in many networks, such as knowledge graphs, it becomes sensitive to noise and outliers.

Maximizing the value of can lead to unsatisfactory results.

의 중요도를 변경하도록 가중치를 부여할 수 있다. 라벨 가중치(label weight)

는

에 따라서

의 중요도를 부여하는 값이다.To solve this problem, for each level transition observation using the maximum weighted likelihood estimation (MWLE) technique

Weight can be assigned to change the importance of. Label weight

The

according to

This value gives the importance of.

는 다음의 수학식 10과 같이 정의된다.Weighted log-likelihood function

Is defined as Equation 10 below.

는

의 값을 최대화한다.Defined by the following equation (11)

The

Maximize the value of

는

인 라벨 전이 관찰의 수를 의미하며, 다음의 수학식 12로 나타낼 수 있다.At this time,

The

It means the number of in-label metastasis observations and can be expressed by the following equation (12).

은 괄호 안의 조건이 맞다면 1의 값을 갖고, 그렇지 않다면 0의 값을 갖는 함수이다.At this time,

Is a function that has a value of 1 if the condition in parentheses is met, and a value of 0 otherwise.

는

및

에 의해 결정됨을 알 수 있다.like this,

The

And

It can be seen that is determined by.

의 값을 1로 놓으면,

는 단지 주어진 레벨 전이 관찰들에 의해서만 결정된다. 또한, 만약

의 값을 높게 설정한다면, 서퍼의 라벨이

로 변경될 확률이 높아진다. 따라서, 최적의 성능을 낼 수 있도록 하는 적절한

의 값을 설정할 필요가 있다.If, all

If you set the value of to 1,

Is only determined by the given level transition observations. Also, if

If you set the value of, the surfer's label

The probability of being changed increases. Therefore, it is appropriate to achieve optimal performance.

It is necessary to set the value of.

Hereinafter, with reference to FIG. 6, a process of learning a rule for changing a label of a surfer, that is, a process of calculating a label transition probability matrix will be described in order.

6 is a flowchart for explaining a process of learning a rule in which a label of a surfer is changed as a surfer moves in a graph labeled with a trunk according to an embodiment, and includes specific steps of step 301 of FIG. 3. .

및 라벨 가중치

를 입력받는다. 이때, 인접 행렬

는 그래프의 간선에 부여된 라벨이 반영된 인접 행렬로서 라벨드 인접 행렬(labeled adjacency matrix)라고도 한다.In step 601, the controller 120 neighbors the matrix reflecting the label assigned to the edge of the graph.

And label weights

Input. At this time, the adjacency matrix

Is an adjacency matrix that reflects the label assigned to the edge of the graph and is also called a labeled adjacency matrix.

를 추출하는 과정을 설명하면 다음과 같다.Labeled adjacency matrix from graph

The process of extracting is as follows.

라고 하고,

가

세트의 정점들과

세트의 간선들로 이루어진다고 하면,

은 간선 라벨들의 세트를 의미한다. 예를 들어

은

으로 표현되고, 이때

는 k번째 라벨이고,

는 간선 라벨들의 개수이다.A graph labeled on the edge

And

end

Set vertices

Speaking of set edges,

Means a set of edge labels. For example

silver

Is expressed as

Is the k-th label,

Is the number of edge labels.

및

가 각각

에 포함되는 정점이고,

가

에서

로 연결되는 간선이라고 하면,

의 라벨은

이다. (

,

,

)

And

Each

Is included in the vertex,

end

in

Speaking of the trunk leading to,

The label of

to be. (

,

,

)

의 라벨드 인접 행렬

는,

로부터

로 연결되는 간선이 있다면 원소

의 값이

이고,

로부터

로 연결되는 간선이 없다면 원소

의 값이 0이 되는 희소 행렬(sparse matrix)이다.At this time,

Labeled adjacency matrix of

Is,

from

If there is an edge leading to the element

The value of

ego,

from

Elements without edges leading to

It is a sparse matrix where the value of is 0.

에 대응되는 그래프로부터 모든 라벨 전이 삼각형(label transitive triangle)을 추출하여 열거한다.In step 602, the controller 120 uses a triangle enumeration algorithm, and the adjacency matrix input in step 601.

Extract all label transitive triangles from the graph corresponding to and list them.

를 기반으로 MWLE를 적용하여 각각의 간선 라벨

에 대한 라벨 전이 확률

를 산출한다.In step 603, the controller 120 extracts the extracted label transition triangle and label weight

Each edge label by applying MWLE based on

Probability of label transition for

Calculate

에 대한 인접 행렬

를 구한 뒤 정규화하는 과정이 필요하다. 이하에서는 도 7을 참조하여 인접 행렬을 정규화하는 과정에 대해서 설명한다.Meanwhile, in order to obtain a score matrix using the LRWR technique, each edge label

Adjacency matrix for

The process of obtaining and normalizing is necessary. Hereinafter, a process of normalizing an adjacency matrix will be described with reference to FIG. 7.

7 is a flowchart illustrating a process of normalizing an adjacent matrix in which a label of a graph is reflected according to an embodiment.

를 입력받는다. 라벨드 인접 행렬 를 구하는 방법은 앞서 도 6의 601 단계에서 설명한 바와 같다.Referring to FIG. 7, in step 701, the controller 120 neighbors the matrix reflecting the label assigned to the edge of the graph.

Input. The method of obtaining the labeled adjacency matrix is as described in step 601 of FIG. 6 above.

에 대해서 준-인접 행렬(labeled semi-adjacency matrix)

를 산출한다. 이때 준-인접 행렬

는, 간선

의 라벨이

이면 원소

의 값이 1이고, 그 외의 경우에는

의 값이 0인 행렬이다. (

는 행렬

의 (u, v)번째 원소)In step 702, the control unit 120 labels each

For labeled semi-adjacency matrix

Calculate The quasi-adjacent matrix

The trunk

The label of

Back side element

The value of is 1, otherwise

Is a matrix whose value is 0. (

The matrix

Of (u, v)th element)

를 산출한다. 이때 연결 차수 행렬

는, 원소

의 값이 정점

로부터 나가는 간선의 개수, 즉 정점 u의 연결 차수(out-degree)인 행렬이다. 연결 차수 행렬

는 다음의 수학식 13 및 수학식 14를 통해 구할 수 있다.In step 703, the controller 120 connects an out-degree diagonal matrix from the labeled adjacency matrix.

Calculate Here, the connection order matrix

Silver, element

The value of vertex

It is a matrix that is the number of edges going out from, that is, the out-degree of the vertex u. Connection order matrix

Can be obtained through Equation 13 and Equation 14 below.

를 이용하여, 각 간선 라벨

에 대한 준-행-정규화 행렬(labeled semi-row-normalized matrix)

를 산출한다. 제어부(120)는 다음의 수학식 15에 연결 차수 행렬

및 준-인접 행렬

를 대입함으로써

를 구할 수 있다.In step 704, the controller 120 connects the connection order matrix.

Using, each trunk label

Labeled semi-row-normalized matrix for

Calculate The control unit 120 is connected to the following equation (15) order matrix

And quasi-adjacent matrix

By substituting

Can be obtained.

에 대한 정규화된 인접 행렬(준-행-정규화 행렬)인

와, 라벨 전이 확률 행렬

가 산출되었다면, 반복적 기법을 이용함으로써 점수 행렬

을 구할 수 있다.Label each edge according to the procedures described above.

Normalized adjacency matrix for (quasi-row-normalization matrix)

Wow, label transition probability matrix

If is calculated, score matrix by using iterative technique

Can be obtained.

을 산출하는 과정에 대해서 설명한다.Hereinafter, a score matrix through an iterative technique will be described with reference to FIG. 8.

The process of calculating is explained.

에 대해서 정규화된 인접 행렬

및 라벨 전이 확률 행렬

, 시작 정점

, 재시작 확률

, 및 수렴 문턱값

를 입력받는다.Referring to FIG. 8, in step 801, the controller 120 labels each trunk.

Adjacency matrix normalized to

And label transition probability matrix

Start vertex

, Restart probability

, And convergence threshold

Input.

를 반영하여 시작 행렬

를 생성하고,

와 같이 점수 행렬을 시작 행렬로 초기화한다. 시작 행렬

는

번째 원소만 1이고, 나머지 원소는 모두 0인 행렬이다. 이때,

는 시작 정점에 대응되는 인덱스 값이고,

는 더미 라벨(dummy label)

를 가리키는 인덱스 값이다.In step 802, the control unit 120 inputs the starting vertex

Reflecting the starting matrix

Create

Initialize the score matrix as Starting matrix

The

The first element is 1, and the rest of the elements are all zero. At this time,

Is the index value corresponding to the starting vertex,

Is a dummy label

Is an index value.

, 라벨 전이 확률 행렬

및 시작 행렬

를 다음의 수학식 16에 대입하고 반복 계산을 수행한다.In step 803, the controller 120 normalizes the adjacency matrix.

, Label transition probability matrix

And starting matrix

Is substituted into the following equation (16) and iterative calculation is performed.

의 원소

는 t번 계산 반복시 라벨

에 대한 시작 정점

와 정점

사이의 점수이며, 다르게 설명하면 랜덤 워크를 t번 반복했을 때 시작 정점

를 출발한 서퍼가 정점

에 도달하여 라벨

를 가질 확률이다.At this time, t denotes the number of times iterative calculation is performed. In addition,

Element of

Is the label when repeating t calculations

Start vertex for

And vertex

This is the score between, and in other words, the starting vertex when the random walk is repeated t times.

The surfer who departed has peaked

Label by reaching

Is the probability of having

와

간 차이의 크기가 수렴 문턱값

보다 작아질 때까지 반복 계산을 수행하고, 차이가

보다 작아지면 점수 행렬이 수렴한 것으로 판단하고 805 단계로 진행하여 수렴된 점수 행렬을 출력한다.In step 804, the controller 120 determines whether the score matrix converges through iterative calculation. In other words, the control unit 120

Wow

The magnitude of the difference between them is the convergence threshold

Iterative calculations are performed until smaller, and the difference is

If it becomes smaller, it is determined that the score matrix has converged, and the process proceeds to step 805 to output the converged score matrix.

The term'~unit' used in the above embodiments means software or hardware components such as a field programmable gate array (FPGA) or an ASIC, and'~unit' performs certain roles. However,'~ wealth' is not limited to software or hardware. The'~ unit' may be configured to be in an addressable storage medium or may be configured to reproduce one or more processors. Thus, as an example,'~ unit' refers to components such as software components, object-oriented software components, class components and task components, processes, functions, attributes, and procedures. , Subroutines, segments of program patent code, drivers, firmware, microcode, circuitry, data, database, data structures, tables, arrays, and variables.

The functions provided within the components and'~units' may be combined into a smaller number of components and'~units', or separated from additional components and'~units'.

In addition, the components and'~ unit' may be implemented to play one or more CPUs in the device or secure multimedia card.

The method of measuring the relevance between vertices of the multi-trunk label graph according to the embodiment described with reference to FIGS. 3 to 8 is also implemented in the form of a computer-readable medium that stores instructions and data executable by a computer. Can be. At this time, instructions and data may be stored in the form of program code, and when executed by a processor, a predetermined program module may be generated to perform a predetermined operation. In addition, the computer-readable medium can be any available medium that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable medium may be a computer recording medium, which is volatile and non-volatile implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. Volatile, removable and non-removable media. For example, the computer recording medium may be a magnetic storage medium such as HDD and SSD, an optical recording medium such as CD, DVD and Blu-ray disk, or a memory included in a server accessible through a network.

Also, a method of measuring the relevance between vertices of a multi-trunk label graph according to the embodiment described with reference to FIGS. 3 to 8 may be implemented as a computer program (or computer program product) including instructions executable by a computer. have. The computer program includes programmable machine instructions processed by a processor and may be implemented in a high-level programming language, object-oriented programming language, assembly language, or machine language. . In addition, the computer program may be recorded on a tangible computer-readable recording medium (eg, memory, hard disk, magnetic/optical medium, or solid-state drive (SSD)).

Accordingly, a method of measuring the relevance between vertices of the multi-trunk label graph according to the embodiments described with reference to FIGS. 3 to 8 may be implemented by executing a computer program as described above by a computing device. The computing device may include at least some of a processor, a memory, a storage device, a high-speed interface connected to the memory and a high-speed expansion port, and a low-speed interface connected to the low-speed bus and the storage device. Each of these components is connected to each other using various buses, and can be mounted on a common motherboard or mounted in other suitable ways.

Here, the processor is capable of processing instructions within the computing device, such as to display graphical information for providing a graphical user interface (GUI) on an external input or output device, such as a display connected to a high-speed interface. Examples are commands stored in memory or storage devices. In other embodiments, multiple processors and/or multiple buses may be used in conjunction with multiple memories and memory types as appropriate. Also, the processor may be implemented as a chipset formed by chips including a plurality of independent analog and/or digital processors.

Memory also stores information within computing devices. In one example, the memory may consist of volatile memory units or a collection thereof. As another example, the memory may consist of non-volatile memory units or a collection thereof. The memory may also be other forms of computer-readable media, such as magnetic or optical disks.

And the storage device can provide a large storage space for the computing device. The storage device may be a computer-readable medium or a configuration including such a medium, and may include, for example, devices within a storage area network (SAN) or other configurations, and may include floppy disk devices, hard disk devices, optical disk devices, Or a tape device, flash memory, or other similar semiconductor memory device or device array.

The above-described embodiments are for illustration only, and those having ordinary knowledge in the technical field to which the above-described embodiments belong can easily be modified into other specific forms without changing the technical idea or essential features of the above-described embodiments. You will understand. Therefore, it should be understood that the above-described embodiments are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope to be protected through the present specification is indicated by the claims, which will be described later, rather than the detailed description, and should be interpreted to include all modified or modified forms derived from the meaning and scope of the claims and equivalent concepts thereof. .

100: relevance measuring device 110: communication unit
120: control unit 130: storage unit

Claims (10)

In the graph labeled the relationship between vertices on the edge, in the method of measuring the relevance between vertices,
Learning a rule in which a label of a surfer changes as a surfer moves between vertices included in the graph;
Normalizing the adjacency matrix of the graph for each edge label; And
And using the result of learning the rule and the normalized adjacency matrix, and calculating a score matrix through an iterative technique. According to claim 1,
Learning the rules for changing the label of the surfer,
Receiving an adjacent matrix and a label weight reflecting the label assigned to the edge of the graph;
Extracting and enumerating all label transition triangles from the graph; And
And calculating a label transition probability matrix for each edge label by applying a maximum weighted likelihood estimation based on the label transition triangle and the label weight. According to claim 2,
The normalizing step,
Receiving an adjacent matrix reflecting a label assigned to the edge of the graph;
Calculating a quasi-adjacent matrix for each edge label;
Calculating a concatenation matrix from the adjacency matrix; And
And calculating a quasi-row-normalized matrix for each edge label using the concatenation order matrix. According to claim 3,
The step of calculating a score matrix through the iterative technique,
Receiving a quasi-row-normalization matrix and a label transition probability matrix, a starting vertex, a restart probability and a convergence threshold for each edge label;
Generating a starting matrix according to the starting vertex, and initializing the score matrix with the starting matrix;
Repeatedly performing a score matrix calculation using the quasi-row-normalized matrix, the label transition probability matrix, and the starting matrix;
Determining whether the score matrix converges based on the convergence threshold value; And
And if the score matrix converges, outputting a converged score matrix. A computer-readable recording medium on which a program for performing the method according to claim 1 is recorded. A computer program stored in a medium for performing the method according to claim 1, performed by a relevance measurement device. In the graph labeled the relationship between the vertices on the edge, in the apparatus for measuring the relevance between vertices,
A communication unit for receiving information on a graph that is a target of relevance measurement;
A storage unit for storing a program for measuring the degree of relevance between vertices of the graph; And
It includes a control unit for measuring the relationship between the vertices of the graph by executing the program,
The control unit learns a rule in which the label of a surfer changes as a surfer moves between vertices included in the graph, normalizes the adjacency matrix of the graph for each edge label, and adjusts the rule. An apparatus using the learned result and the normalized adjacency matrix and calculating a score matrix through an iterative technique. The method of claim 7,
The control unit,
When the adjacency matrix and label weight reflecting the label assigned to the edge of the graph are received, all the label transition triangles are extracted and enumerated, and the maximum weighted likelihood estimation is applied based on the label transition triangle and the label weight. And a label transition probability matrix for each edge label. The method of claim 8,
The control unit,
When an adjacency matrix reflecting the label assigned to the edge of the graph is received, a quasi-adjacent matrix is calculated for each edge label, and then a concatenation order matrix is calculated from the adjacency matrix, respectively using the concatenation order matrix. And calculating a quasi-row-normalized matrix for the edge labels of the. The method of claim 9,
The control unit,
When inputting a quasi-row-normalization matrix and a label transition probability matrix, a starting vertex, a restart probability and a convergence threshold for each edge label, a starting matrix is generated according to the starting vertex, and the score matrix is the starting matrix. After initializing to, the score matrix calculation is repeatedly performed using the quasi-row-normalization matrix, the label transition probability matrix, and the starting matrix to converge the score if the score matrix converges based on the convergence threshold. Device for outputting a matrix.

Images